The neural basis of behavior is believed to consist of sequential patterns of neural activity in the relevant brain regions. Behavioral flexibility also requires neural circuit mechanisms that support dynamic control of sequential activity. However, mechanisms to control and reconfigure sequential activity have received little attention. Here, we show that recurrently connected networks with heterogeneous connectivity and a smooth spatial in-degree landscape (which may arise due to asymmetric neuron morphologies) provide a robust mechanism to evoke and control sequential activity. By modulating the synaptic strength of only a few neurons in local neighborhoods, we uncovered high-impact locations that can start, stop, extend, gate, and redirect sequences. Interestingly, high-impact locations coincide with mid in-degree regions. We demonstrate that these motifs can flexibly reconfigure sequential activity, and hence, provide a framework for fast and flexible computations on behavioral timescales, while the individual parts of the pathways remain rigid and reliable.

Mesocorticostriatal dopamine projections are crucial for value learning, motivational control, and cognitive functions. However, while dopamine's role in value learning as reward-prediction-error (RPE) has been much understood, precise roles in motivational control and cognitive functions remain more elusive. Computationally, this corresponds to that while the operation of mesostriatal dopamine could be minimally described by simple reinforcement learning (RL) models with one-dimensional reward/RPE and fixed state representation, (1) how reward-specific motivational control can be achieved through heterogeneous dopamine responses, and (2) how sophisticated cortical state representation can be formed through mesocortical dopamine, cannot be captured by such simple models. To address both of these at once, we combined recent models for each of them: the "Reward Bases (RB)," which achieved reward-specific motivational control through multidimensional RPE (but with fixed cortical representation), and the "online value-recurrent-neutral-network (OVRNN)," which achieved state representation learning through training of RNN by RPE (but of one-dimensional). We show the combined model can achieve both functions simultaneously via double "feedback alignments" of the cortical and striatal downstream connections to the mesocorticostriatal dopamine projections. Crucially, cortical inhibition-dominance is a key for successful learning. Excessive excitation leads to aberrant persistent activity, which disrupts the alignments and impairs reward-specific motivational control and credit assignment. This implies how negative and positive symptoms of schizophrenia could emerge from excitation/inhibition imbalance, and we show how our model could explain altered brain activations in patients. Our model thus provides an integrated computational account for dopamine's functions, with implications on how its dysfunctions link to schizophrenia.

Despite the diversity and volume of brain data acquired and advanced AI-based algorithms to analyze them, brain features are rarely used in clinics for diagnosis and prognosis. Here we argue that the field continues to rely on cohort comparisons to seek biomarkers, despite the well-established degeneracy of brain features. Using a thought experiment (Brain Swap), we show that more data and more powerful algorithms will not be sufficient to identify biomarkers of brain diseases. We argue that instead of comparing patient versus healthy controls using single data type, we should use multimodal (e.g. brain activity, neurotransmitters, neuromodulators, brain imaging) and longitudinal brain data to guide the grouping before defining multidimensional biomarkers for brain diseases.

The striatum comprises a network characterized by a highly shared feedforward inhibition (FFI) mediated by fast-spiking interneurons (FSI), which constitute only 1% of the striatal population. We investigated the dynamical consequences of this extensively shared FFI beyond inducing synchrony in a local striatal microcircuit. Our findings reveal that increased FFI sharing enhances the across-trial variability of striatal responses, activity of medium spiny neurons (MSNs), to cortical inputs, and endows the striatal network with the capacity to modulate output correlations in a bidirectional manner. Specifically, weakly shared cortical inputs become more correlated, whereas strongly shared cortical inputs are decorrelated in the presence of FSIs. These dynamic modulatory effects on MSNs, in turn, substantially alter the spiking statistics of downstream neurons in the globus pallidus, regarding across-trial variability and burstiness.

Networks in the brain operate in an excitation-inhibition (EI) balanced state. Altered EI balance underlies aberrant dynamics and impaired information processing. Given its importance, it is crucial to establish non-invasive measures of the EI balance. Previous studies have suggested that relative EI balance can be inferred from the spectrum of the population signals such as Local Field Potentials (LFP), Electroencephalogram (EEG) and Magnetoencephalography (MEG). This idea exploits the fact that in most cases excitatory and inhibitory synapses have quite different time constants. However, it is not clear to what extent spectral slope of population activity is related to the network parameters that define the EI balance e.g. excitatory and inhibitory conductance. To address this question we simulated two different types of recurrent networks and measured spectral slope for a wide range of parameters. Our results show that the slope of the spectrum cannot predict the ratio of excitatory and inhibitory synaptic conductance. Only in a small set of simulations a change in the spectral slope was consistent with the corresponding change in the synaptic weights or inputs to the network. Thus, our results show that we should be careful in interpreting the change in the slope of the population activity spectrum.

Neural activity sequences are ubiquitous in the brain and play pivotal roles in functions such as long-term memory formation and motor control. While conditions for storing and reactivating individual sequences have been thoroughly characterized, it remains unclear how multiple sequences may interact when activated simultaneously in recurrent neural networks. This question is especially relevant for weak sequences, comprised of fewer neurons, competing against strong sequences. Using a non-linear rate -based and a spiking model with discrete, pre-configured assemblies, we demonstrate that weak sequences can compensate for their competitive disadvantage either by increasing excitatory connections between subsequent assemblies or by cooperating with other co-active sequences. Further, our models suggest that such cooperation can negatively affect sequence speed unless subsequently active assemblies are paired. Our analysis characterizes the conditions for successful sequence progression in isolated, competing, and cooperating assembly sequences, and identifies the distinct contributions of recurrent and feed-forward projections. This proof-of-principle study shows how even disadvantaged sequences can be prioritized for reactivation, a process which has recently been implicated in hippocampal memory processing.

Representation of external and internal states in the brain plays a critical role in enabling suitable behavior. Recent studies suggest that state representation and state value can be simultaneously learned through Temporal-Difference-Reinforcement-Learning (TDRL) and Backpropagation-Through-Time (BPTT) in recurrent neural networks (RNNs) and their readout. However, neural implementation of such learning remains unclear as BPTT requires offline update using transported downstream weights, which is suggested to be biologically implausible. We demonstrate that simple online training of RNNs using TD reward prediction error and random feedback, without additional memory or eligibility trace, can still learn the structure of tasks with cue–reward delay and timing variability. This is because TD learning itself is a solution for temporal credit assignment, and feedback alignment, a mechanism originally proposed for supervised learning, enables gradient approximation without weight transport. Furthermore, we show that biologically constraining downstream weights and random feedback to be non-negative not only preserves learning but may even enhance it because the non-negative constraint ensures loose alignment—allowing the downstream and feedback weights to roughly align from the beginning. These results provide insights into the neural mechanisms underlying the learning of state representation and value, highlighting the potential of random feedback and biological constraints.

Models of visual salience detection rely on center-surround interactions, yet it remains unclear how these computations are distributed across retinal, cortical, and subcortical circuits due to their overlapping contributions. Here, we reveal a de novo collicular mechanism for surround suppression by combining patterned optogenetics with whole-cell recordings from individual neurons in the mouse superficial superior colliculus (SCs). Center zones were defined by monosynaptic input from channelrhodopsin-expressing retinal ganglion cells in collicular midbrain slices. Surround network optoactivation suppressed center responses compared to center-only input. This suppression is excitatory in origin, arising from the withdrawal of center excitation via surround-driven inhibition of local recurrent excitatory circuits, as demonstrated by cell-type-specific trans-synaptic tracing and computational modeling. These findings identify a local circuit mechanism for saliency computation in the SCs, independent of cortical input.

Researchers have begun to use the synergy of psychiatry and neuroscience to identify biomarkers that can be used to diagnose mental health disorders, predict their progression, and forecast treatment efficacy. However, biomarkers have achieved limited success to date, potentially due to a narrow focus on specific aspects of brain signals. This highlights a critical need for methodologies that can fully exploit the potential of neuroscience to transform psychiatric practice. In recent years, there has been emerging evidence of the ubiquity and importance of moment-to-moment neural variability for brain function. Single-neuron recordings and computational models have demonstrated the significance of variability even at the microscopic level. Concurrently, studies involving healthy humans using neuroimaging recording techniques have strongly indicated that neural variability, which in the past was dismissed as undesirable noise, is an important substrate for cognition. Given the cognitive disruption seen in several psychiatric disorders, neural variability is a promising biomarker in this context, and careful consideration of design choices is necessary to advance the field. In this review, we provide an overview of the significance and substrates of neural variability across different recording modalities and spatial scales. We also review the existing evidence that supports its relevance in the study of psychiatric disorders. Finally, we advocate for future research to investigate neural variability within disorder-relevant, task-based paradigms and longitudinal designs. Supported by computational models of brain activity, this framework holds the potential for advancing precision psychiatry in a powerful and experimentally feasible manner.